jevbooks

← All projects

jev-visual

hr98w/jev-visual

Educational MLX vision-language inference with shared image context and direct candidate scoring.

Scoring & Ranking82%Runner-up: Evaluation & BenchmarkingMany questions, one request

What it is

A small runnable Python project for learning vision-language model inference on Apple Silicon. It uses Qwen3.5-0.8B with MLX to answer multiple questions about one image, with a local browser UI, CLI and HTTP API.

How it uses Jev

It explores a Jev-like inference pattern: reuse shared multimodal context, fork the cache, batch question suffixes, and score candidate outputs directly from model logits instead of autoregressive structured generation. Candidate probabilities are relative to supplied options and are not correctness estimates or Jev's calibration.

Primitives:choicescore

Technique worth stealing

Reusing shared multimodal context and directly scoring candidate outputs from logits instead of autoregressive generation.

Try it

Install uv, create a Python 3.13 venv, install requirements, run jev-visual-download, then start uvicorn on 127.0.0.1:8788.

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricyes0.68
  • Shows a System One patternyes0.76
  • Handles uncertaintyno0.04
  • Measuredno0.06
  • Runnableyes0.73
  • Worth recommendingno0.33
  • Model replicayes0.87
  • Problem scopescore on a 0–2 scale0.15
  • About Jevyes0.76

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.